Experiment: Increase UNEMBEDDING_LR from 0.01 to 0.015 to test whether a higher learning rate for the output layer improves validation bpb.

Campaign

• Campaign: optimizer-tuning

Hypothesis

Increase UNEMBEDDING_LR from 0.01 to 0.015 to test whether a higher learning rate for the output layer improves validation bpb.

Parent Context

• Parent master hash: 935fdbf9f4ae8a5ef5bcb76552acea2bc5801965
• Master val_bpb at dispatch: 0.962777
• Worker id: worker-0
• Worktree: /Users/ben/code/open-autolab/.runtime/worktrees/unembedding-lr-015

Single Variable

UNEMBEDDING_LR = 0.015 (was 0.01)

Expected Upside

Higher learning rate for the output layer may allow faster adaptation of the final projection, potentially improving validation bpb.

Duplicate Check

No prior experiments have tested UNEMBEDDING_LR changes against this master.

Runtime

• Log path: /Users/ben/code/open-autolab/research/live/unembedding-lr-015.log
• Launcher: uv run scripts/opencode_worker.py run unembedding-lr-015

Allowed Edit Scope

• train.py only

Run Plan

• Refresh master with uv run scripts/refresh_master.py --fetch-dag
• Run uv run scripts/hf_job.py preflight
• Run uv run scripts/hf_job.py launch --mode experiment
• Stream logs to the reserved path
• Parse uv run scripts/parse_metric.py /Users/ben/code/open-autolab/research/live/unembedding-lr-015.log

Result

• HF Job: 69cc18d8942f980bf425a641
• val_bpb: 0.970805
• Master val_bpb: 0.962777
• Delta: +0.008028 (degraded)
• training_seconds: 300.1
• total_seconds: 431.5
• peak_vram_mb: 33609.6
• mfu_percent: 44.14
• total_tokens_M: 306.6
• num_steps: 2339
• Submitted: no
• Interpretation: Increasing UNEMBEDDING_LR from 0.01 to 0.015 significantly degraded validation bpb. The output layer is sensitive to learning rate increases.
• Failure mode: Significant regression (+0.008028 bpb).

Memory-Keeper Handoff

• One short note for research/notes.md: Added experiment entry with FAILED status.
• Any do-not-repeat update: Added to Optimizer Tuning section.